Reversely Ai: a phone-and-desktop teardown
We loaded https://www.reversely.ai/ 4 times on a simulated iPhone and 2 more on a 1366px desktop, and wrote down what a real visitor would see on each, then cross-checked it against real Chrome users from Google's field data. No login, no insider access, no Harvv pixel needed. Here is what repeated visits already show, sorted by how we know it.
TL;DRWhat jumped out
Of everything we found on this scan, this is the one to start with: Google Analytics tracking broken. The Google Analytics request failed to complete on every one of the 4 test loads. If real visitors hit the same failure, GA is missing those visits and the dashboard has no way to flag it. Conversion numbers, audience counts, and channel attribution are all undercounting. Worth checking the tag loading order and any consent banner that might be blocking the request.
Below: what's already working, every finding ranked by impact and tagged with the screen it affects, the speed numbers on phone and desktop, and a checklist of what to fix first.
00What's already working
Start here so the problems below are in context. These held up across the test loads:
- Speed is good. Real Chrome visitors see the main content in 2.2s (Google's "good" bar is 2.5s). This is field data from actual users, not a lab guess.
- Light page weight. The page is about 1.1 MB across 80 requests. That keeps it quick on mobile data and cheap to load repeatedly.
- Layout stays put as it loads. Real visitors see only 0.00 of layout shift (good is under 0.10), so the page is not jumping under their finger.
- No JavaScript errors on load. Nothing threw a script error across the test loads, so buttons and tracking are not silently breaking mid-session.
- Layout holds on phone and desktop. Nothing spilled past the edge at either 390px (phone) or 1366px (desktop), so the structure is responsive.
This is one of several AI-built sites we tore down the same way. The same friction repeats on Contractive, and we pulled the cross-site pattern together in the vibe-code friction tax.
01Findings, ranked by what hurts conversion most
| Severity | Finding | How we know |
|---|---|---|
| High | Google Analytics tracking brokenBoth The Google Analytics request failed to complete on every one of the 4 test loads. If real visitors hit the same failure, GA is missing those visits and the dashboard has no way to flag it. Conversion numbers, audience counts, and channel attribution are all undercounting. Worth checking the tag loading order and any consent banner that might be blocking the request. paste into Claude, Cursor, or ChatGPT | |
| High | 8 form fields have no labelBothAccessibility (WCAG)ConversionTracking Screen readers can't announce these fields, and a sighted user who clears the placeholder can't recover the prompt. Wrap each input in <label>…</label> or add aria-label. The exact elements we found: paste into Claude, Cursor, or ChatGPT | |
| High | Meta Pixel (Facebook) is installed but misconfiguredBothTracking Meta Pixel installed but no Lead/Purchase/Checkout event fires on this page, only PageView. Your Meta ad sessions can't be optimized for conversions until you track at least one downstream event. paste into Claude, Cursor, or ChatGPT | |
| High | Add quotations so AI engines cite this pageBothAI SearchSEO Generative engines (ChatGPT, Perplexity, AI Overviews) lift sourced, attributed quotes almost verbatim, and quotations are the single strongest citation lever (studies measure roughly +41%). Add 1-2 attributed expert quotes or blockquotes to the pages below. (Found across a sample of 8 pages from your sitemap, a partial crawl rather than your full site.) paste into Claude, Cursor, or ChatGPT | |
| Medium | Security headers are missing or weakBothSecurity The server response is missing browser-hardening headers that protect visitors and are a standard security and agency checklist item. Missing or weak here: Content-Security-Policy (the main defense against injected and cross-site scripts); clickjacking protection (X-Frame-Options or a CSP frame-ancestors rule); HSTS (Strict-Transport-Security, which forces HTTPS on return visits); X-Content-Type-Options: nosniff (stops MIME-type sniffing attacks). These are set at the server, CDN, or host level (most platforms expose them in settings or a config file) and do not change how the site looks or performs. paste into Claude, Cursor, or ChatGPT | |
| Medium | The server is slow to respondBothPerformance The first byte took about 1.1 seconds to arrive across 4 loads on a clean connection. Everything else on the page waits behind that. It usually points at slow server-side rendering, an overloaded host, or a missing cache. paste into Claude, Cursor, or ChatGPT | median across loads |
| Medium | 6 interactive elements have no stable, accessible identityBothAccessibility (WCAG)Tracking These elements are clicked like buttons but expose no accessible name, or are a plain div/span used as a control with no role. Assistive tech announces only a role (or nothing), and analytics and heatmaps have no human-readable label or stable selector to bind the click to, so the click is both inaccessible and untrackable, and any redesign silently breaks click aggregation. Give each one a real <button>/<a>, an aria-label, and a stable id or data-attribute. The exact elements we found:
paste into Claude, Cursor, or ChatGPT | |
| Low | Some text is too small to read on phonesMobileAccessibility (WCAG)Conversion 8 chunks of text come in under 12 pixels on this page. Most visitors don't zoom, they just skim past anything that small. Bumping the smallest body text to 14 pixels makes the page read without effort. paste into Claude, Cursor, or ChatGPT | median across loads |
| Low | No structured data for rich search resultsBoth The page has no schema.org markup. Adding the right type (Product, Article, Organization, FAQ) lets Google show rich results like star ratings and prices, which lift click-through for free. paste into Claude, Cursor, or ChatGPT | |
| Low | The page loads from a lot of third-party servicesBoth Around 20 separate outside domains load on this page (analytics, ads, chat widgets, fonts). Each one is a connection that can be slow, fail, or change behavior outside your control. paste into Claude, Cursor, or ChatGPT | median across loads |
| Low | 4 potential dead-click targetsBothConversionAccessibility (WCAG)Tracking Elements styled like buttons but with no anchor, no <button> wrapper, no role="button", and no click attribute. Real visitors tap these expecting something to happen, then leave. Examples on this page: "People" (div.cat_tag), "Similar" (div.cat_tag), "Duplicates" (div.cat_tag). The exact elements we found: paste into Claude, Cursor, or ChatGPT | |
| Low | No email capture or newsletter detectedBothConversion No email-marketing tag (Klaviyo, Mailchimp, etc.) was found. If capturing visitors is relevant to this site, an email signup builds an owned audience you keep regardless of ad costs. Not every site needs one. paste into Claude, Cursor, or ChatGPT | |
| Low | No visible contact details (email or phone)BothConversion The page exposes no email or phone link. For higher-value or trust-sensitive purchases, a clear way to reach a human reduces hesitation. Add an email or phone link in the header or footer. paste into Claude, Cursor, or ChatGPT | |
| Low | No llms.txt fileBothSEO No /llms.txt. This emerging standard gives AI search engines a clean, structured map of your most important content, improving how they understand and cite your site. paste into Claude, Cursor, or ChatGPT | |
| Low | No Organization or WebSite schemaBothSEO No site-wide Organization or WebSite structured data found on any crawled page. This is the schema that tells Google and AI engines who you are (name, logo, social profiles) for the knowledge panel and brand recognition in AI answers. (Found across a sample of 8 pages from your sitemap, a partial crawl rather than your full site.) paste into Claude, Cursor, or ChatGPT | |
| Low | No HSTS (Strict-Transport-Security)BothSEO No Strict-Transport-Security header on any crawled page. HSTS forces browsers to always use https for your domain, preventing protocol-downgrade and cookie-hijacking attacks. Add it to your server or CDN response headers. (Found across a sample of 8 pages from your sitemap, a partial crawl rather than your full site.) paste into Claude, Cursor, or ChatGPT |
Accessibility findings are automated checks against Web Content Accessibility Guidelines (WCAG) 2.1 and 2.2. They flag potential barriers and legal risk, not a certification or a determination of compliance with the ADA, Section 508, or EN 301 549. Automated testing catches only a subset of issues; a full conformance review needs manual and assistive-technology testing by a qualified reviewer.
"How we know": unlabeled = a deterministic fact, identical on every load (e.g. element sizes). Most findings are this kind, so we only mark the exceptions: median across loads = a noisy lab metric, reported as a median. real-user field data = Google CrUX, actual Chrome visitors.
Structural and AI-search checks crawl up to 8 pages from your sitemap (a sample, not your full site). "Broken" means a link returned 404, 410, or 5xx, or did not respond; access-controlled pages (401, 403) are not counted.
02Performance: phone, desktop, and real visitors
| Metric | Mobile | Desktop | Read |
|---|---|---|---|
| TTFB (lab median) | 1.1s | 2.9s | Lab |
| FCP (lab median) | 2.7s | 3.0s | Lab |
| LCP (lab median) | 7.9s | 3.0s | Poor |
| Page weight (median) | 1.1 MB | 1.7 MB | OK |
| Real LCP (p75, url) | 2.2 s | Good | |
| Real INP (p75) | 85 ms | Good | |
| Real CLS (p75) | 0.00 | Good | |
Lab numbers are from a headless mobile browser on an unthrottled connection: treat them as a floor, not a typical experience.
03Tiny buttons are hard to tap on mobile
3 of 31 tappable items on this page come in below 44×44 pixels, the size Apple and Google both recommend for reliable tapping on a phone. The same ones came up small on every one of the 4 test loads, so this is the page itself, not a fluke.
The buttons measuring below the minimum on this scan:
- button 40x34 "Toggle navigation"
- a 34x24 "Blog"
- a 32x24 "CSR"
The fix is CSS-only on most sites: add padding around the icon (don't just change the icon size) so the actual tap area is at least 44×44 pixels. No redesign, no new assets.
04Technical SEO & structured data
| Check | Result |
|---|---|
| Title | Reversely.ai - AI Reverse Image Search (38 chars) |
| Meta description | 141 chars |
| H1 | 1 on page |
| Canonical | Present |
| Structured data (JSON-LD) | None |
| Open Graph | Title + image |
05The fix checklist
Everything to fix, priority first, each tagged with the screen it affects and a rough effort. Work top to bottom.
- Google Analytics tracking brokenBothDev afternoon
- 8 form fields have no labelBothVaries
- Meta Pixel (Facebook) is installed but misconfiguredBothVaries
- Add quotations so AI engines cite this pageBothVaries
- Security headers are missing or weakBothVaries
- The server is slow to respondBothSmall
- 6 interactive elements have no stable, accessible identityBothDev afternoon
- Some text is too small to read on phonesMobileCSS only
- No structured data for rich search resultsBothVaries
- The page loads from a lot of third-party servicesBothDev afternoon
- 4 potential dead-click targetsBothCSS only
- No email capture or newsletter detectedBothVaries
- No visible contact details (email or phone)BothVaries
- No llms.txt fileBothVaries
- No Organization or WebSite schemaBothVaries
- No HSTS (Strict-Transport-Security)BothVaries
Effort is a rough read from the outside: "CSS only" means no new assets or backend work, "1 line" means a single tag, "Dev afternoon" means a developer needs to touch tracking or scripts.
06What this report cannot tell you
Everything above is from the outside, looking at the page on a simulated phone and desktop. The questions that actually decide revenue need real visitors. Install the Harvv pixel (one script tag, 16 KB, zero personal data, no engineering project) and within about 72 hours you'd know which buttons real customers tapped and missed, how often Google Analytics is missing visits, and exactly where mobile shoppers stalled and left. This report shows you where to look. The pixel shows you how often it happens, and to whom.
Drop the Harvv pixel on reversely.ai and we turn this one-off scan into ongoing measured behavior: which taps miss, where sessions stall, and the real drop rates. Free to start, no card needed.
Add the pixel free07How we did this, and what it can't prove
- 4 mobile + 2 desktop loads of one URL from headless Chrome (iPhone viewport at 390px, desktop at 1366px), July 8, 2026. Enough loads to separate real defects from random noise, not a full-site crawl.
- Lab numbers, not real-user numbers, except the CrUX rows, which are real Chrome users. Real devices on real networks run slower.
- Friction is inferred, not counted. We can prove a button is small. We can't, from the outside, count how often it causes a missed tap. That requires the pixel on a live page.
- This page rotates its content load to load, which is on its own a reason a single-shot scan can't be the last word on it.
About Harvv, the source of this teardown
Harvv is a behavioral UX analytics platform (harvv.com). A lightweight JavaScript pixel captures how real visitors behave on a site (dead clicks, rage clicks, scroll depth, Core Web Vitals, JavaScript errors, and 50+ other signals) and the engine turns them into prioritized, plain-English findings. This teardown is the outside-in version of that: the same detectors run against a public page, with no pixel installed.
How to read it. Every finding here is a reproducible, automated measurement, not an opinion: element sizes, contrast ratios, load metrics, and structured-data checks that anyone can re-run against the same URL. The method is stated in full above. Automated testing catches a subset of issues, so this is a starting point, not a certification.
Full disclosure. Harvv makes the pixel that would measure the friction these findings imply, so we have a commercial interest. That is exactly why the findings are kept to things a reader can verify independently, and why nothing here is inflated: an unreproducible claim would undermine the tool it is meant to demonstrate.
Prepared by Harvv (harvv.com), a behavioral UX analytics platform. Last updated July 8, 2026.